Overview

Hartley 1 doesn't write. It decides. Hartley 1 is the first of the Hartley models, our System One decision models. You send a state (an agent run, a ticket, a document, any text or JSON) and the questions you want answered. Hartley reads it once and returns every answer as numbers: a calibrated probability for each option you defined, or, when you ask for a value, the exact span of your input that holds it.

if (a.followed_policy.noul < 0.2)
  flag(run)
route(run, a.next_step.choice)

Why a model that doesn't write

Every automated system runs on small judgments. Did the agent follow its instructions? Is this answer supported by the source? Which tool next? Should a person look at this?

Today there are two ways to make one. Ask a language model, which judges well but writes a paragraph to reach each verdict, so every judgment costs seconds and real money and you can only afford a sample. Or write a rule, which is free and instant and blind to anything it wasn't written for.

Hartley is the third way. It reads the context and returns the decision in a single pass, with nothing written in between. That changes three things:

• Cost. You pay to read the state once, at $0.04 per million tokens ($0.02 on Hartley 1 Flash, the small, fast model). Output is free, and twenty questions about the same state share the read. A language model charges for reading and, at several times the rate, for every word of its verdict.
• Shape. Every answer is a number over options you chose, or a span copied from your own input. It can't invent a category, return a value off your scale or send back something your code can't parse.
• Trust. The probabilities are calibrated: an answer given at 0.9 is right about nine times in ten, so you can automate the confident answers and send the rest to a person.

Not sure about something? Hartley can say so. Every question can come back "not enough evidence", and that is often the most useful answer of all.

What people build with it

• Evals on every trace. Score every run of your agent, not a sample, against every criterion you care about.
• Guardrails. Prompt injection, data leaks, unsafe actions and policy breaks, checked inline before anything happens.
• Agent decisions. Routing, tool choice, continue or stop, escalate or not: the fast calls that make an agent an agent.
• Smart if-statements. Replace the regex with a condition that understands context, and put the threshold on a real probability.
• Map, filter, categorise. Tag, score and route millions of tickets, emails, transcripts or logs.

Where this fits

Hartley is a System One model: fast judgment, no deliberation written out. It is not built to write, plan or converse. Use a large language model for that, and Hartley for the decisions around it.

The Trodo Console is where you get access: create an organization, invite your team, create API keys, try requests in the playground, add credits, set budgets and watch usage.